AI agent security needs controls across the stack

AI security should be treated as an engineering discipline with defined requirements, enforceable controls, named owners and evidence that protections work. As AI agents reason, use tools and adapt to data, organizations need to apply established security principles, including identity, access control, exposure limits and verification, to new operating conditions.

Protection must span the full agent stack. Models provide capabilities, harnesses organize context and workflows, and runtime environments determine what actions can execute. An agent allowed to update a customer record should not automatically be able to export customer data, and any request for additional access must be authorized outside the agent itself. Network policies, sandboxing, limited credentials and protected logs help enforce those boundaries and support investigations.

NVIDIA OpenShell is presented as an open source secure runtime that enforces policies outside an agent’s reach, while Open Secure AI Alliance partners are building related governance, verification and policy tools. Examples include Cisco DefenseClaw, JFrog integrations for scanning and verifying agent skills, CrowdStrike SafeMind for repeated attack simulations, Palo Alto Networks Prisma AIRS for continuous red teaming, Capital One VulnHunter for code security and ReversingLabs Spectra Assure for detecting malware and tampering in software packages.

Deployment decisions should rely on tests that show controls can block credential misuse, unauthorized data transfers, permission changes and attempts to interfere with monitoring. Failures should become repeatable tests, and shared research, tools and evidence can help defenders improve protections as agent capabilities advance.